United States · counties
What the county measures describe
Many county measures rise and fall together, so they repeat each other. This page asks how many truly separate patterns they hold. Looking at levels (how counties differ at one point in time), 77 county measures behave like about 6.0 separate things. The first of these patterns accounts for 37% of all the differences between counties. It moves with income (in percentage terms), at r = -0.90. Here r shows how closely two measures rise and fall together, from −1 (opposite) to +1 (in step), and 0 means no link. So much of what sets one county apart from another is how rich it is, measured in many ways. Change over time is different. Over 2014–2019, 43 measures behave like about 24 things, and no single pattern accounts for more than 11%.
Levels, 2023
77 measures across the 2,381 counties that report all of them. The table calls each pattern an axis. Together, 10axes account for 80% of the differences between counties. Counts are left out, since they show how big a county is rather than what it is like. Where several measures track one quantity, such as seven elections’ vote shares, only the latest is kept. The axes are not named by hand. Each is described by the measures that line up most strongly with it, at its high and low ends. It is also described by how closely it moves with income and with population density (both in percentage terms). “Share” is how much of the differences that axis accounts for.
| Axis | Share | With income | With density | High end | Low end |
|---|---|---|---|---|---|
| 1 | 37% | -0.90 | -0.30 | Independent-living disability; Mobility disability; Coronary heart disease | Median household income (model-based); Returns reporting capital gains; Share with a bachelor's degree or higher |
| 2 | 13% | -0.15 | -0.55 | Share White, not Hispanic; Cancer other than skin cancer; Homeownership rate | Share never married; Median gross rent; Returns claiming charitable contributions |
| 3 | 6% | +0.09 | +0.26 | Net migration per 100 people staying; Arrivals per 100 people staying; Mean travel time to work | Effective property tax rate; Adult income rank of children raised poor here; Binge drinking |
| 4 | 6% | +0.10 | -0.33 | People per tax return; Share Hispanic or Latino, any race; Adults under 65 without health insurance | Routine checkup in the past year; Cholesterol screening; Current asthma |
| 5 | 5% | -0.17 | -0.31 | Share aged 65 and over; Vacancy rate; Median age | Routine checkup in the past year; Taking blood pressure medication; People per tax return |
| 6 | 4% | -0.03 | +0.03 | Depression; Current asthma; Membership of volunteering groups | Share in the same house a year ago; Taking blood pressure medication; Homeownership rate |
Change, 2014–2019
The same breakdown, done on each county’s change: 43 measures across 1,678 counties. Here 23 axes are needed for 80%. Change spreads in so many directions that grouping measures is useful here in a way it is not for levels. In levels, a group mostly just finds income again. And every comparison of two measures on this site already adjusts for income.
| Axis | Share | With income | With density | High end | Low end |
|---|---|---|---|---|---|
| 1 | 11% | +0.22 | -0.23 | Returns claiming the earned income credit; Unemployment rate; Poverty rate (model-based) | Mean adjusted gross income per return; Per capita personal income; Median age |
| 2 | 7% | +0.27 | +0.22 | Median home value; Median gross rent; Median household income (model-based) | Effective property tax rate; Premature mortality; Share White, not Hispanic |
| 3 | 6% | +0.46 | +0.10 | People per tax return; Share now married; Poverty rate (model-based) | Returns claiming charitable contributions; Share never married; Share Hispanic or Latino, any race |
| 4 | 5% | -0.19 | -0.04 | Share now married; Years of potential life lost; Premature mortality | Share never married; Share divorced; Children in single-parent households |
| 5 | 5% | +0.29 | +0.14 | Net business formation rate; Business opening rate; Poverty rate (model-based) | Business closing rate; Jail admission rate; Jail population rate |
| 6 | 4% | +0.31 | +0.29 | Injury deaths; Years of potential life lost; Premature mortality | Net business formation rate; Jail population rate; Returns claiming charitable contributions |
Combined measures for change
When two publishers measure the same county’s change, they agree far less than they do on its level. A change is the difference between two estimates, and each estimate has some random error, so the errors pile up. Where several publishers measure one quantity independently, their average is more reliable than any single one. “Independently” is the key condition. A measure estimated from another one, or adjusted to the same population estimates, cannot confirm it, so those are left out. A candidate is kept only if its members agree about change at ≥0.2, on a scale where 1 is perfect agreement and 0 is none. “Reliability” shows how trustworthy the average is; higher is better, up to 1.
| Measure | From | Members agree | Reliability | Findings that held |
|---|---|---|---|---|
| Income growth | Per capita income; Per capita personal income; Mean adjusted gross income per return | +0.37 | +0.64 | 3 of 6 a lower rate than single measures |
| Population growth | Population; People covered by a tax return | +0.79 | +0.88 | 31 of 41 |
“Findings that held” checks whether results repeat. It takes pairs of independent measures that moved together over 2014–2019, comparing counties only with other counties in the same state. The link had to be unlikely to be chance (q < 0.05, a test that allows for checking many pairs at once). The pair also had to be measurable again over the second period. It then counts how many held: they moved the same way again, and again in a way unlikely to be chance (q < 0.05). Pairs with a combined measure held 72% of the time (34 of 47). Pairs without one held 52% (265 of 510).
How change is measured
- 2014–2019: the main test: each Census survey figure averages five years, and these two do not overlap; before the pandemic.
- 2019–2023: a repeat check: its two five-year survey figures both include 2019, and it covers the COVID years.
- Money and counts are compared as growth in percentage terms. Rates are compared as the change in percentage points.
- Each county’s change is compared with its own state’s average change. The calculation also allows for the fact that counties in the same state tend to resemble each other. So a regional shock that hits both measures at once does not show up as a link.
- Two measures from the American Community Survey share their sampling error in change. (Sampling error is the random error from surveying only some people.) Both come from the same people answering the survey. Those pairs are labelled and kept out of the count of findings. So are pairs built from the same numbers or the same model, which are bound to match.
- Values carried forward from earlier years to fill gaps in the 2023 snapshot are not used. Only years in which a measure was actually observed count.